Active Participant

—

Predictions generated from wearable PPG-derived model inputs. Dexcom values are used only as retrospective ground truth.

Model Estimate

—

Model Confidence

—

Dexcom Ground Truth

Hidden

Used only for validation

Subject Agreement

—

Dexcom Ground-Truth Timeline

Retrospective glucose trajectory synchronized with each model prediction window.

Model Probability Vector

Hybrid probability output for the selected window.

Hypoglycemia —

In Range —

Hyperglycemia —

Uncertainty Gate

—

Model Benchmark

Chronological personalized evaluation: 70% train / 15% validation / 15% test.

Hybrid selected

Subject Glycemic Distribution

Hypo

—

In Range

—

Hyper

—

Session Range

Peak

—

Nadir

—

Prediction Audit Log

Compare model inference with independent CGM ground truth.

Timestamp	Subject	CGM	Ground Truth	Prediction	Probabilities	Confidence	Result

Research Context

Input

Five-minute wearable PPG windows and engineered physiological signal features.

Architecture

CatBoost gradient-boosted trees combined with a Temporal Convolutional Network through validation-selected probability fusion.

Validation

Dexcom CGM acts as retrospective ground truth and is not supplied as an input to the prediction model.

GlycemicPlus is an experimental research prototype. Current performance is not sufficient for clinical use. It should not be used for diagnosis, insulin dosing, or treatment decisions.